Embed Size (px)
Citation preview
Zero-Carbon Cloud: A Volatile Resource for High-Performance Computing
Andrew A. ChienUniversity of Chicago and
Argonne National [email protected]
Rich WolskiUniversity of California, Santa Barbara
Fan YangUniversity of Chicago
Abstract—The growing deployment of renewable powergeneration creates a growing opportunity in “stranded power”,power that is generated a close to zero-cost, but not usable bythe power grid. We propose to use this stranded power tocreate a “zero carbon” high-performance computing resource,exploiting the batch computing model to exploit the volatilepower efficiently.
Keywords-Data Center; Stranded Power; Green computing;sustainable computing;
I. INTRODUCTION
Data and digital assets are rapidly becoming the fun-damental building blocks for society. Commerce, govern-ment, education, science, and even social interaction aredigital endeavors that are consummated and optimized withinformation. However, the byproducts of computing arebecoming endemic in society, with its carbon footprintsignificant in mankind’s total carbon emission (computingcontributes over 2% of global carbon emissions [?]). Likewith all ubiquitous societal technologies, the managementof the resources consumed becomes a societal issue aswell. Expanding computing in conventional fashion to meetfuture ambitions would further increase the damaging carbonemissions of ICT. Our ambition is to create a new source ofcomputing which has dramatically lower carbon footprint,called “Zero-Carbon Cloud”.
So how can we create this transformative breakthrough?The key ideas are to 1) the harness the excess of intermittentpower created by the shift to renewables, and 2) exploit light-weight data center infrastructure to radically reduce otherelements of total-cost of ownership (TCO). Our concept,“Zero-Carbon Cloud” combines these to create limitless,low-cost computing with different volatility characteristicsthan the ones exhibited by computing systems today. Wepropose ZCCloud that exploits volatile renewable powergeneration the non-interactive (batch) model of most high-performance computing to create a low-cost powerful HPCcomputing resource. Note that ZCCloud is a pure renewable-based computing services, a radical contrast to greeningefforts [?] that purchase a balancing average of renewablepower.
It is also worth noting that renewable’s such as solar andwind variable output makes them challenging for integrationinto a reliable power grid. Such grids have been designed
and engineered for controllable generation at fixed locations– and highly optimized for connectivity and cost on thatbasis. The shifting quantity (and consequently location)of power generation by renewables increases transmissionrequirements due to congestion and greater distance fromgeneration to consumption. This poses serious technical andeconomic challenges for the power grid [?]. The widespreadadoption of ambitious Renewable Portfolio Standards (RPS)that set goals for rapid growth in the fraction of renewablepower that utilities must employ, the variability challengeis tremendous and growing. The dynamic range of suchresources exceeds 50% of peak load today, and may increaseto 100% within 15 years [?]. To the grid, Zero-Carbon Cloudis an example of a dispatchable load, that both creates ondemand a high-value service, but and a new volatile formof computing.
New intellectual concepts for volatile cloud systems andapplications, including systems, service-level agreements,prediction, and scheduling and marketplace are needed torealize this vision.
II. RENEWABLE OPPORTUNITY
Growing concerns about the impact on climate and envi-ronment of carbon emissions resulting from burning of fossilfuels [?], [?], [?], [?], have led to large-scale deploymentof renewable sources of electricity generation. By far thefastest growing types, and those projected to address asignificant fraction (¿10%) of demand are wind and solar[?], [?]. The variability and non-dispatchable nature ofthese renewable sources, combined with low incrementalgeneration cost creates significant challenges for power-griddesign and management [?], [?], [?], [?]. At present, whengeneration exceeds demand and the excess power exceedsthe grid storage’s limited abilities, it is simply discarded atthe source – it is “stranded power.” Power grids call this lossof excess power “curtailment” or “down dispatching”. It isthis opportunity that we propose to exploit with ZCCloud.
Numerous power grids (Independent System Operators– ISO’s) around the world have stranded power, Figure 1reports data from the Midcontinent Independent System Op-erator (MISO), showing total generation, total wind power,and total power “curtailed/down-dispatched” for a recent twoand a half year period. Despite improved grid connectivity
Figure 1. Stranded Power in the MISO Power Grid, June 2011-October2013 [?]
and management has, and MISO’s economic dispatchingmarket still suffers from a few percent waste, an extraor-dinary amount of power – 1.6 TWh in 2014. Comparablelevels prevail for ERCOT (wind) and CAISO (solar andwind), and numerous regions in Europe (Denmark, Germany,Ireland, Italy) [?]. In all of these power grids, the fractionof renewables is expected to increase by 100% or morein the next decade, creating even greater challenges tothe maintenance of power-grid balance, and more strandedpower [?], [?].
In the MISO region, wind power has significant penetra-tion today with smaller states such as Iowa and Minnesota¿20%, and larger states such as Illinois and Michigan at5%, but all of thes states have adopted Renewable PowerStandards (RPS) goals to double this percentage by 2025.[?].
Figure 2. Total Cost of Ownership, based on [?], for low-cost serverand partially filled data center scenarios. ZCCloud can substantially reducepower and data center costs, accounting for 63-70% of TCO.
III. ZCCLOUD APPROACH
The basic approach of Zero-Carbon Cloud (ZCCloud)is to exploit recent technological advances in InformationTechnology (e.g. cloud computing, data center automation,system-focused analytics, etc.) to leverage “stranded” powerin renewable energy settings. The result is a cloud computing
Figure 3. Zero Carbon Cloud Reduced Costs: Containers (C) and StrandedPower (S)
capability lower fixed cost (lower physical plant), lowervariable costs (lower power costs), and ultimately loweroverall TCO for delivered computing.
Published data suggests data centers with more than 50%physical plant and nearly 20% server costs with the sumaccounting for about 75% of the TCO [?], [?]. The primaryelement of the remaining 25% is electricity as illustrated inFigure 2.
ZCCloud can reduce these costs in two ways.1) Using containerized server facilties, sited at renewable
generation sites, ZCCloud eliminates the need forpurpose-built buildings to house the infrastricture andpower distribution.
2) Exploiting stranded power, the cost of power can bereduced well below even the wholesale prices paid bylarge data centers , perhaps ten-fold.
We outline a case study of traditional data center inFigure 3, along with a projection for achievable ZCCloudcost. Together these improvements suggest a system thatcould provide computing 2-fold cheaper by exploiting adecentralized architecture, and as much as 3-fold cheaperif stranded power is exploited.
ZCCloud provides an energy sink for curtailed renewablepower that is capable of producing useful computationalwork. That is, the curtailment is simply converted to compu-tation and storage capabilities (albeit with different volatilityand availability cycles than traditional systems) rather thanbeing discarded as extra, unconsumable power. Once theexcess power becomes computation, however, modern datacenter efficiency technologies can be used to maximize theutilization of this power. For example, heat reclamationtechniques commonly used today [?] in many top-end datacenters to maximize energy efficiency become relevant. Thatis, the curtailed power that is converted to computing andstorage can be made to do so with increasing efficiencies us-ing the technological advances that are improving datacenterefficiencies today and in the future.
IV. DESIGN AND SCALING ZCCLOUD
Superficially, wind or solar power may seem to be unus-able for computing due to their intermittent availability. In
Figure 4. Scaling from Small: 2-container, 2.2 Petaflops, 0.59MW, Medium: 4-container, 4.4 PF both can be power by a single turbine. Extreme:42-container, 45.5 PF, 12.4MW, 5% of turbines in a wind farm.
Figure 5. MISO Wind Generation for One Week.
fact the rates of available wind power change on long timescales (several days or weeks) or as short as a few hours.Solar power is more predictable, varying in a similar dailycycle approximately matched to increases in traditional de-mand (daytime higher, nighttime lower). We detail a sketchZCcloud system design, and then discuss intermittence.
Design of ZCcloud Building Blocks: ZCCloud usesconvention building blocks that achieve high densities ofcomputing per rack and per container. The computing nodesare connected with low-latency, high-speed 10-gigabit Eth-ernet switches. We assume the containers have a PowerUsage Effectiveness (PUE) of 2.0, which means the non-compute facilities, e.g. cooling system etc. has the samepower consumption as compute nodes. This conservative andpublished numbers for commercial container-based productsand hyperscalers such as Google are now below 1.25 andas low as 1.19. To enable real-time response even whenstranded power is unavailable, ZCCloud also deploy analways-on frontend server for each container, the powerconsumption of which is nontrivial comparing to the totalcontainer power. The resulting power and computing densityis summarized in Table I.
Scaling ZCCloud: Significant computing facilities are ofmodest scale compared to modern wind farms. Our smallscale system of 2.2 petaflops and medium scale systemof 4.4 petaflops require below 1.2MW and thus could be
easily placed beneath a single modern 2MW turbine – thedominant size being deployed in commercial farms today[?] (see Figure 4) Our Large system is still modest in size.A 42-container system with 45.5 petaflops capacity wouldrequire 2 containers at the bases of 21 wind turbines –a small fraction of a modern commercial wind farm. Forexample, the Twin Groves farm [?] includes 240 turbines,each 1.65MW for peak generating capacity of 398MW .Even a 3% curtailment can power significant computingcapacity (our Large system of 45.5 PF is only 12.4MW ).From our 2.2 petaflop ZCCloud building block, replicationto achieve 100-petaflop systems is straightforward. One suchwould cover only 18.9% of a large wind farm such as TwinGroves, and there are dozens of such facilities in the MISOregion, so total capacity exceeding exaflops is possible.
ZCCloud offers intermittent and variable capacity basedon the availability of stranded power. Volatility tied towind power will support continuous availability from hours(overnight) to days, due to the change in weather patterns(see Figure 5). Volatility tied to solar stranded power appearsto be likely tied to regular daily and weekly cycles, butmay involve shorter periods. Commercial uses and marketsfor volatile computing resources exist. Large-scale cloudprovides, like Amazon AWS, offer “spot instances” – rentalsat a bid “spot price” that are terminated when the market (orperhaps the provider) decides that they should be reclaimed.
Description Performance PowerNode Dual sockets, Intel E7-8890 v3 CPUs 4,838 GFLOPS 0.66 kWRack 14 nodes per rack 67.732 TFLOPS 9.24 kWContainer 16 racks per container 1.083 PFLOPS 295.68 kW
Table IZCCLOUD COMPUTING CONTAINER SKETCH DESIGN
In short, they can be revoked at any time, yet are deemeduseful by a large user community [?]. Unlike these othervolatile cloud computng rentals, however, ZCCloud volatil-ity results from the fluctuations in available power and notmarket or other commercial forces. Thus it is possible to of-fer more reliable minimum guarantees of service (comparedto current spot-market offerings) in the form of Service LevelAgreements (SLAs). In short, usage will resemble Amazon’sspot-instance facility but adding a guarantee of minimumtime to “eviction” and subsequent spot-instance termination.Extensions such as Amazon Spot Fleet’s, combining sets ofthese systems (or their virtualization in instances) also makesense.
We plan to enhance the capabilities of ZCCloud throughthe construction and operation of a series of software andhardware prototypes. These prototypes will demonstrate theeconomic benefits of ZCCloud, create a volatile computingresource - demonstration to application users -, and en-able advanced research to improve service-level agreements(SLA’s) to increase the value of the delivered computingservices.
V. INITIAL EVALUATION
To assess the utility of a Zero-Carbon Cloud HPC comput-ing resource, we have performed a series of simulation usingover 12 months of job traces from the Argonne LeadershipComputing Facility’s Mira system [?], considering a numberof different stranded power scenarios. Using a system withtwo times the hardware resources, but only intermittentpower of 8 hours/day (33% duty factor), our preliminaryresults show:
1) >30% of the jobs experience comparable or betterturnaround time
2) the largest jobs experience improved turnaround time3) Simple rules can identify which jobs will benefit,
making ZCCloud useful as a complementary resourceto traditional HPC platforms.
In short, we are encouraged that ZCCloud is a promis-ing approach to create a new class of HPC computingresources – supporting new capabilities and promsing forcost-effectiveness.
VI. SUMMARY AND DISCUSSION
We have described the Zero-Carbon cloud concept, de-scribing the basic elements of exploiting stranded powerand light-weight physical infrastructure. Our initial design
and evaluation suggests the approach is promising – withthe potential to dramatically reduce computing costs, andcreate a complementary capability to traditional approaches.We look forward to exploring the ZCCloud concept moredeeply in the future, including open challenges and potentialdirections, including 1) rigorous simulation studies showingthe benefits and exploring the huge configuration space ofZCCloud systems, 2) a detailed design and demonstrationof the ZCCloud system, 3) exploring the addition of limitedenergy storage, 4) exploring opportunities of geographicdistribution, including job migration, and many more.
ACKNOWLEDGEMENTS
This work was supported by in part the Office of Ad-vanced Scientific Computing Research, Office of Science,U.S. Department of Energy, under Award DE-SC0008603and Contract DE-AC02-06CH11357, as well as the NationalScience Foundation under Awards CNS-1405959, STCI-0751315, CNS-0905237, and CNS-1218808. The authorsalso gratefully acknowledge generous support from HP,Keysight, Huawei, Nvidia, and the Seymour Goodman Foun-dation.
